I agree that you could do more with less computing power if you started with 1000x1000 than with 30x30
However, the software engineering and algorithm design time would be much greater for the 1000x1000 case, IMO ben > -----Original Message----- > From: [EMAIL PROTECTED] [mailto:[EMAIL PROTECTED] > Behalf Of Brad Wyble > Sent: Thursday, December 23, 2004 11:18 AM > To: [email protected] > Subject: RE: [agi] RE: AGIRI fundraising drive... > > > > In each case *some* customization of the generic processes will > be needed in > > order to make them achieve these tasks. But I believe that the > amount of > > customization needed is much greater for the 1000x1000 case than for the > > 30x30 case. > > Depends what you are trying to do with the information. > > Obviously 30x30 is far less information than 1000x1000, this makes it > easier to process exhaustively, but on the other hand, it's harder to do > something interesting with that impoverished data stream. You may find > that upstream of low-level visual processing, it's actually a greater > challenge to do something of interest with a very low res information > stream. > > An upper level process would prefer to have a 1000x1000 representation > pared down to 30x30 (trivial), have the same transformations applied that > you would have done to the 30x30, but then be able to "zoom in" where > deemed necessary, getting more detail. > > In short, you could probably do more with less computing power if you > started with a 1000x1000 input stream than with a 30x30, because you have > the flexibility of examining either resolution. > > This is why our brains are inundated with attentional mechanisms that > operated within and across every sensory modality. Simple processing > where sufficient, deeper processing where necessary gets you more > bang for > you buck. > > > -Brad > > > > > > > > I suspect that: > > > > * in the 30x30 case we can get away with some minor specializations and > > customizations and then let the general cognition tools do the > work, whereas > > > > * in the 1000x1000 case we badly need a more complex hierarchical > > architecture in which the lower levels are dealt with by very > specialized > > stuff, and the higher levels are more like the > > mildly-specialized-general-cognition algorithms that are *all* > you need in > > the 30x30 case... > > > > -- Ben > > > > > >>> If a 30x30 pixel grid is what you mean by content-richness, then I do > >>> intend > >>> for Novamente to deal with *this* kind of content-richness in > the fairly > >>> near future (how near depending on the achievement of > relevant funding, > >>> blah > >>> blah blah). I believe that this level of richness doesn't require the > >>> kind > >>> of complex, specialized pre-filtering that human-eye-level richness > >>> requires... > >>> > >> Urgh, the adjective applies to what goes on *inside*, not what comes in > >> from the outside. If you just create 30x30 PixelNodes and let > the generic > >> processes do the rest, you don't have content-richness (or if > you do, you > >> have to learn *all* of it). Its a question of what kind of build-in > >> support you have for stuff like the things I mentioned (resolving > >> ambiguity, object completion, invariance under > transformations, temporal > >> patterns).. And yeah, I know novamente can recognize temoral patterns > >> ;->... > >> > >> Moshe > > > > > > ------- > > To unsubscribe, change your address, or temporarily deactivate > your subscription, > > please go to http://v2.listbox.com/member/[EMAIL PROTECTED] > > > > ------- > To unsubscribe, change your address, or temporarily deactivate > your subscription, > please go to http://v2.listbox.com/member/[EMAIL PROTECTED] > > ------- To unsubscribe, change your address, or temporarily deactivate your subscription, please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
